Zitron's argument, built from the "construction in progress" lines in hyperscaler earnings reports, is that a large share of the GPUs Nvidia has sold are sitting in warehouses or in data centers without power. He is explicit that it's his own estimate from incomplete disclosure — Microsoft doesn't report the figure at all. The point he draws from it is that chip sales measure willingness to speculate on AI, not demand for it.
So local opposition, maybe I'm wrong. This is one where I feel this, but maybe I will be proven wrong in history. I don't think it's local opposition. I think that slowing things down and any slowdown with these things makes them more expensive. Any of them. Like it's really, it's bad news. Every day is more interest added onto the loan. It's not great. I think it's that these things are the most ambitious infrastructure projects of all time and they're trying to do a bazillion of them. I really think I think it's a much more simple thing of these things are tough to do. Every single one has, depending on the ground you're building on, there might be a combination of clay, might be not as stable, depending on the weather, depending on the humidity, heat dissipation is different with each one. The power grid in a particular place you're building it might be a little bit decrepit. In Port Washington, Wisconsin, Oracle was forced to have a, I think they're being forced to do a $7 billion bond requiring $100 million a year in payments just to secure that if Oracle doesn't. So actually, taking a step back, in a lot of these situations, taxpayers would bear the brunt of the costs in the event that the power wasn't built. Because even if it's a state-owned utility, and even if the company's paying for the utility to do it or somehow compensating them, if the company dies or doesn't build the data center, the taxpayer will be the one that covers the distance. In Wisconsin, they said, screw that, we're not accepting that. You have to have a security bond that says if Oracle pulls out of this, WE Energy will still have the money to build it. And the power companies are kind of even state-owned ones, kind of scurrilous when it comes to this stuff. They screw the taxpayer wherever they need to. And all they want to do is make money. So Oracle is already having to do that. I think the local opposition that matters is this power thing, though. It's the people who are, it's when there is a worry that they will not complete these data centers, they will say, well, it's going to be because of the power costs. It's not that the leftover incomplete data center would be bad, but the incomplete power infrastructure, you have to finish that. Depending on where it is in the construction phase, you actually, for safety reasons, have to finish that. Anyway, Bernstein, who's an analyst group, put out a report this month that said 35%, only 35% of all data centers announced ever get built. I think that's high. I think it's 10% or less. But the point is, while it might be local opposition, well, it might hurt them. The real problem is these things are so bloody difficult to build. They're so complex. They're so expensive. And they require specialist talent. They require electric grade steel that's in the shortage. They require transformers. And when they don't have the transformers in, they use behind the meter power with these gas turbines that blow up, well, not blow up. They run it. They kind of wind themselves down because you're not meant to run these things as often as they do. So you have these kind of cobbled together projects that are drawing a city's worth of power into this tiny space. Every possible variable is there for things to go wrong. And now things are going wrong. So people are saying, well, I don't know. But the big thing is, is if Oracle does any more deals, I don't know if they're all done with the specific, because they've got these data center SPVs, which they connect to the projects. I don't know if they're done raising those. I hope they are for their sake. Because if Oracle has to do another one of these SPVs, they're going to ask for a vial of Larry's blood. They're going to ask for one of the large Ellison sons. They're going to keep him in a room until Larry pays for the data center. And honestly, based on their relationship, I'm not sure Larry gets him out. But putting all that aside, these things are, Oracle is now, I wouldn't say persona non grata to the debt markets. God know. They'll take, look at what they still lend to Masayoshi's son. What I think is more likely now is they're going to turn the screws on them. And the big problem that Oracle has right now, a little bit of a technical thing, but it's much simpler. Oracle's credit rating right now is one step above junk. Now, when you're investment grade, you get bought by pension funds, investment funds, hedge funds, and they buy you just because they go, this is an investment grade company. Within our diverse portfolio, we have bonds with this company. If Oracle gets downgraded to junk level, all of their bonds will get sold off. And their bonds are already... So when you sell off a bond, technically a bond's like $100. When you sell off a bond, its price goes down, but the effective yield of the bond goes up just because you're paying less than $100 for it. So if it was 6%, it's now effectively 8%. And what that really means is the market is asking. If Oracle borrows again, which they will have to, they are asking Oracle to be priced as junk. If that officially happens, if SP officially downgrades them, Oracle enters a world of hurt and becomes something called a fallen angel, which is a term for when an investment grade becomes a junk, has all their bonds sold off, and it just makes their credit so much more expensive. Now, this is this has happened, I think, with Ford in the past. I think it nearly happened to Boeing, but I can't remember. Point is, it doesn't happen often, much like the force majeure doesn't happen often. And I repeat something I've been saying for months, which is Oracle, people think that I'm going to bake my bread with OpenAI dying. I think they might be right. I think people know me for that. I actually think Oracle dying is far more realistic. Well, I think OpenAI dying as well, but Oracle is right now, even though they are not the principal connected to this debt, even though you technically couldn't come after Oracle's assets for this. Oracle runs on debt, has run for decades on debt, on borrowing, endless borrowing. If you make Oracle's way of life untenable, which is borrowing, I don't know what Oracle does. And also, Oracle is not growing anymore. The revenue's been flat for 15 years if you adjust for inflation. The only thing growing is AI. How does the AI grow without the data centers, Larry? How does it work? It's crazy that this is not all over the finance press and the business press and the tech press. I mean, they're covering the sporce majeure thing, but the brittle state of Oracle and its relationship with OpenAI is going to be the story of the AI bubble.
And that does lead me nicely on to the next bit, which is that Oracle is a primary buyer of NVIDIA GPUs. And if the development in New Mexico ends up cancelled, they're going to have to have so many spare GPUs they'll kind of be using as paperweights, let alone anything else. I mean, you've done a lot of digging into sort of in your free newsletter this week into the numbers behind how many GPUs there are sitting around. We did a little bit of a React episode, kind of, I guess it was to the story with Microsoft having not brought on as much compute as we thought they did. Since then, what have you found and how many GPUs are being hoarded?
So the number of GPUs is impossible to calculate because everyone obfuscates them. But let's start with the really simple thing, which is I estimate that 200 to 300 billion dollars worth of NVIDIA GPUs are either in warehouses or sitting in unpowered data centers. And it gets back to the thing I was saying with Oracle. Data centers are ambitious projects. They're not coming along very fast. I got to this number because Microsoft doesn't disclose this number, but Google, Amazon, on a yearly basis for Amazon, on a quarterly basis for Google, Meta, SpaceX, and a bunch of other companies have something called Construction in Progress. Now, CIP is a, well, that can refer to other things. Construction in progress refers specifically to stuff that is not being put in service. Now, if you go through their earnings, as I have, because I'm a curious little critter, for the most part, they say 50 to 60% of that is data center hardware. So not construction, but like data center chips, short-lived assets is the euphemism they use. So I think the number was around $374 billion of CIP across SpaceX, Google, Meta, Amazon, Core Weave, Applied Digital, a bunch of others. Forgive me for not having the list memorized fully. However, this list is not actually as large as it is because Microsoft does not disclose construction in progress. Based on my discussion with my source, Microsoft has about $50 billion, $60 billion max worth of AI chips installed and in service. Microsoft has spent over $265 billion on capital expenditures. This leads me to believe that they have about 100, 120 billion or so of chips total. So somewhere between 60 and 80 billion, just uninstalled. Amazon, their number is about 71 billion construction in progress, but that was from the end of 2025. They only update it annually. So that will be much higher. So we are looking at 400,500, I think. And that sounds people are like, people really, the boosters have been pissy online. They're very upset with me because it's a damning thing. Because there's a big difference between everyone has sold this idea of this massive demand for AI. It's insatiable demand for AI. And it's kind of sold on this lie of company buys GPU, company installs GPU, money comes out of GPU. That story is materially different if the GPUs that NVIDIA sells every quarter take two years to put in. Because it means that buying NVIDIA GPUs is not actually proof of anything. NVIDIA's sales are proof that Jensen Huang is the world's most successful FinDom. He just texts Satch, Satcha, you pig, give me $10 billion. And Sag goes, yes, sir. Yes. So, of course, my terminal says that Microsoft's NVIDIA's largest customer. So, that must be how it works. I'm kidding. I don't think it's literally the text-based thing. But it shows that people are just pre-ordering these GPUs, that there's not demand for them based off of, oh, I know that I can turn this into money for sure. It's demand based on fear of missing out and industrial pressure. So, all I don't like, I don't think the majority of Blackwell GPUs, which started selling in 2025, in seriousness, I don't think most of those are installed. I think it the idea that people are talking about Vera Rubin is hilarious to me. It's like talking about flying cars and talking dogs. It's just like, wow, Vera Rubin, we should see that in 2030. And people will say, well, if Microsoft showed they installed one. Yeah, they installed a server in a building. And when you go and look, it's in the middle of a floor, which is not how data centers look. So it's like, yeah, you plug one of these things in, but it's the key jingling. It's a jingle jingle. Oh, look, we've got a Vera Rubin. Oh, it's so good and fast. And everyone goes, yay, everything's happening. Except nothing's happening. Like, it sounds like maybe $150 billion worth of GPUs out of the 535 billion, maybe 500 billion sold, have actually been installed anywhere. And this is bad for NVIDIA. It's bad for Broadcom. Broadcom estimates they're getting, sorry, analysts estimate that Broadcom's going to make about $300 billion selling TPUs to Anthropic and Google and also chips to OpenAI in the next three years. How's that going to, first of all, how are they going to pay for them, of course, but also are those going nowhere too? And what it means is that supply, what people think is happening is NVIDIA sells GPU three months later. It's turned into money and it's sold to this diverse group of customers. And there's all this demand and that's why revenues are going up. What's actually happening is this dribble of capacity is coming online. It's a little dribble and it's immediately being sold to OpenAI and Anthropic. Microsoft, Google, Amazon, they just sell it immediately on whatever it is Meta's doing. It's like having a drunk, it's like having a drunk friend leaving you a seven-minute long voice note. You're like, I made a target, man. It's so good. All right, Mark, I'll call you in the morning. Drink some water. But the big thing is, is that we can no longer take any announcement of capacity seriously. That is the big message. I don't know how there is not an SEC investigation into this. Microsoft has said for the past three quarters that it brought a gigawatt of capacity online. And you, as a person, are meant to think, oh, AI capacity. When Brody over at Bloomberg reported that out of 12 gigawatts, only two gigawatts was actually AI capacity. And by the way, I'm being that curious little critter that I am, I went and I actually read a lot of the earnings going back years. They only recently started mentioning data center capacity at all in fiscal year 25, so sometime in 2024, calendar year. So Microsoft is obviously, when they say capacity, they want you to think AI. They are saying it in the context of AI, but they don't mean AI. Microsoft is not bringing a gigawatt of capacity online. And Aisha Down and I from The Guardian, when Aisha went and asked the question and said, like, hey, this two gigawatt thing, because we went and brought the numbers, they came back and said, they're reading the numbers wrong. They're completely wrong, blah, blah, blah. So obviously just Microsoft talking out their rear end. But I think this is across the industry. I actually don't know if more than a gigawatt of AI capacity came online last year. I don't know if more than a gigawatt came on this year. And the fact we don't know is completely and utterly insane. It is completely deranged. I think the fact that journalists are not banging on their doors, demanding answers, the fact that it took this long for anyone to get off their ass and do something about it and say something and look into this is a joke. It's a failure of modern journalism. And I don't know how we have got to this point where journalism is so weak that we don't ask the simplest and most obvious question of the AI bubble, which is how many AI chips are there? How many of them are working? How many of the chips that NVIDIA has sold are actually in service? Because it sounds to me, and based on everything I've seen, that it's a fraction of the amount sold. And in failing to interrogate this and in failing to actually work this number out and actually quantify what's being sold, journalism has helped sell a massive lie. Because think of it like this. If you are looking at NVIDIA and they're selling all these chips and you think those chips that they sell every quarter are quickly generating revenue, that's a very different story to if they're being warehoused. It means that there's obviously growing demand for AI because you're actually. You're selling more and more chips every year, and more and more chips are being put in service, and more and more chips are making money. But if it's just a dribble, a non-specific dribble is coming online, but more and more are being sold. All you are measuring is how willing people are to speculate on AI. And that's a terrible measure of an industry's success. If anything, it's the antithesis of demand. It's speculation is the root of all evil, to quote Gordon Gecko from Wall Street to Money Never Sleeps. And that is a real quote. It's great, great movie, right up until the end.
I mean, one of the big questions on the topic of big questions that we came away with last week was: if why are people still buying more GPUs if they're just putting them in a warehouse? Have you gained any more insight on that at all?
I think that it's a combination of things. One, I think the hyperscalers are doing just dim-witted industrial planning. They are saying, okay, I'm going to build expazillion gigawatts of G of AI compute. I need to buy this many GPUs. I also think that Jensen Huang of NVIDIA has made it very clear that if they don't do that, he will not sell them to them in the future. I think that there was a Wall Street Journal article about Google's TPUs that had this line in it about how if neoclouds buy other people's chips, they get put in something called Jensen jail, Jensen Huang, CEO of NVIDIA, where they don't get sold GPUs. By the way, this is a massive antitrust problem. Like, this is the FTC should be all over this and aggressively policing it because that is illegal under the Sherman. Like, that's the most that like you are literally doing anyway. You, unless you sign an exclusive contract and have actual terms and compensation, which they do not. So, I think that Jensen Huang has said to people: if you do not buy these all in advance, I will not help you. I will not get you the chips. I cannot guarantee them. And because hyperscalers are cowards, because they are all cowards, every single one of them, they don't use their position as the largest vendors to just go, No, no, Jensen. I will buy these when the data center is ready. And Jensen would go, Oh, okay, I understand. Because that's the thing. Jensen has the power here because, which really started in 2023 when there was the initial shortage of chips. And so, the H100 specifically and H200 eventually, and there was so much demand that they had to put them in armored cars to deliver them, which is kind of a little much. But they, I think, that because of that, kind of like the supply chain shortage with the Suez Canal a few years ago, everyone just bought everything they could. The problem there is: has anyone double-ordered? Has anyone double-ordered from NVIDIA? How can you cancel these GPUs? And also, important legal question: my mate in Fixed Income asked this: Do they exist? If they don't exist, if this is a paper thing, then NVIDIA is probably in real legal trouble. So, I'm going to just assume they're in warehouses in Taiwan, if I had to guess. But that's a lot of chips to warehouse. So, I think we'll eventually find out, but I think it's this should be a five-alarm fire. This should be something that makes everyone quite concerned. I'm not trying to tell people how to feel. I'm just like, everyone is selling the dream of AI demand and GPU sales. People justify the growth of the AI industry. If you say to someone, AI is kind of a con, or AI isn't really as big as people think. They point at NVIDIA and they say NVIDIA sell lots of GPUs. That wouldn't happen if AI wasn't a big deal. Except, it absolutely would, and it literally is. It's not like this is ridiculous. Look at Project Jupiter. Look at Project Abilene. Look at all of these data centers that are taking two, three, four years to build. And even then, do they have the power? Of course, these things aren't being installed. It's laughable. And I've been saying this for like a year. And I keep saying it and it keeps coming true. I think by the time that people go, oh, yeah, Ed was onto something, all of this will be in ruins. And I think that the eventual outcome is going to be, and you can also, there's a depreciation line on the balance sheets of all these companies as well, or income statements, pardon me, where the depreciation is not growing with CapEx. So something only starts depreciating. So you spread out the cost over a few years, only starts happening when they're put in service. Their depreciation is not growing very fast. Google's like six, seven billion. Amazon's got like 20 billion, but they have a ton of different warehouses that it makes sense for. But no one, this should be 20, 30, 40 billion dollars worth of depreciation each. Like it's, there's so much being built. Unless it's not, unless it's just in warehousing. So yeah, one day we're going to find out. I'm going to be right. Everyone's going to say, oh, how obvious. Except it was obvious in 2025. It was obvious today. It'll be obvious next year. It's been obvious the whole time. If you choose to look, no one wants to because it's so gross when you do.
Where does it leave? I mean, GPU bag debt. If I cannot believe that it was sold on the terms or the knowledge or the expectation that there's all these chips just sitting around not being used. And the thing with the depreciation that I kind of get stuck on, which does lead, does connect to this as well, is that it might not start until they're plugged in, but they're getting more obsolete every year with each new release as well, anyway. So does that not depreciate them as well?